ldwang
ldwang
@tridao If there is no causal_conv1d_fn , how does the normal [conv1d](https://github.com/state-spaces/mamba/blob/main/mamba_ssm/modules/mamba_simple.py#L64-L72) perform causally? Thanks https://github.com/state-spaces/mamba/blob/main/mamba_ssm/modules/mamba_simple.py#L168
https://github.com/state-spaces/mamba/issues/108
You can manually update flagai sources. https://github.com/FlagAI-Open/FlagAI/commit/0fc0cc2fb70cbc151119929cb440922802b241d5 Later we plan to make a new release.
I tested, but int4 costs 2 time of FP16. Anything wrong?
> Some zero data caused it! @980202006 could you explain this root cause and what is zero data like, thanks.
> Great work! Great work! Look forward to a script of convert megatron mixtral to hf.
> 同样关注这个能力，是否有这部分能力的强化？对于企业应用，这块能力太重要了 请问国内哪些模型这个能力比较强，谢谢
智源社区的模型文件和huggingface一样，可以按需下载任一即可。
https://modelscope.cn/models/BAAI/AquilaChat2-7B